Your knowledge.
Your intelligence.
Investigate a failing release. Find the current procedure. Recover the context behind a decision. We’re building AI for company work, on infrastructure you control.
Company Brain is in development. Try the local AI available now.
From question to knowledge your team can review.
“Checkout is failing after this release. Have we seen this before?”
This checkout release changed the connection-pool setting [2]. A prior incident recovered after that setting was reverted [1]. The current cause is unconfirmed.
Draft hypothesis The earlier connection-pool failure may have recurred.
“Checkout recovered after the connection-pool setting was reverted.”
- 01InvestigateGather evidence
- 02ReviewCorrect the draft
- 03Approve knowledgeOwner approval required
- 04Reuse in the next investigationOnly approved updates
A planned workflow using a model chosen by your team on infrastructure you own. Examples and review are illustrative; nothing is saved or used for training here.
Engineering investigations are the first workflow in development. Operations and onboarding examples show future possibilities.
What we’re buildingKnow the work.
Keep what you learn.
From engineering investigations to the procedures and decisions that keep a company moving.
Company Brain
In developmentInvestigate a failed release with relevant runbooks, past incidents, and changes together in a reviewable brief.
Owned intelligence
Local runtime available · organization controls plannedRun approved models on machines and infrastructure you control. Choose where the work happens.
Reviewed improvement
PlannedCapture corrections and real outcomes. Test changes before adopting them, with model adaptation when it proves useful.
Start with an engineering investigation, then carry the evidence and open questions into a handoff.
See the first workflowYour teams. Your sources. Your compute.
Start with one chosen machine. Bring more owned capacity into the picture as your needs grow.
Workstation or private server, with additional owned capacity explored experimentally.
- Sources attached
- Owner review
- Scoped reuse planned
Intended for reuse only within an authorized scope, once organization controls are available.
Illustrative concept for organizations of different sizes. Multi-machine model splitting is experimental; this is not live activity.
Free at the core.
Built for company work.
Run Potluck yourself, or help shape the company tools we’re building for operating it across your organization.
Core · $0
Local AI availableFree software for individuals and companies, on your own hardware. The cited investigation and reviewed knowledge workflow is being built into the core.
Company
In developmentPlanned source monitoring, fleet administration, controlled updates, and support for a private deployment.
Your work stays yours.
Our product commitmentBasic security, source traceability, and data portability belong in the core. Company size doesn’t make core use a paid feature.
Core has no software license fee. You supply the hardware and any third-party services you choose. Company pricing and pilot scope are discussed together.
See pricing and compare plansOne 32B model.
A PC and a Mac.
We’ve run Qwen2.5-Coder 32B across an RTX 2080 Ti Linux PC and an M2 Pro Mac. Both held model weights. Both helped generate the answer.
See the configuration and resultsWhere the model actually ran.
- Linux PC · CPU
- 6.23 GiB
- Linux PC · GPU
- 7.58 GiB
- Mac worker
- 1.03 GiB
Measured model-weight placement · Q3_K_M · September 29, 2026
Development builds. Some weights remained on the PC’s CPU. This proves a model ran across both machines; it does not establish a speedup or that neither could run it alone.
A bigger idea.
A little film.
Start with the computers you already own. Then imagine what becomes possible when more people can bring their machines to the same network.
Why we’re building PotluckA 30-second concept film about where we’re headed.
Read the transcript
It starts with an idea. Connect the computers around you. At home. At work. With friends. Across the country. Together, power larger AI models. Bring your ideas to life. Potluck. Bring what you’ve got. Build something bigger.
Your machines
Start hereLocal AI today. Managed model splitting in development.
A trusted circle
In developmentAn invited group sharing capacity.
An open pool
In developmentA wider network of contributing members.
Shared-network research remains on the roadmap. Our immediate product focus is current company knowledge, useful work, and reviewed outcomes.
Follow the roadmapStart with what
you already have.
Try the free local runtime on a workstation you control. Explore owned-machine execution while we build the company investigation and review workflow.
Download the desktop app.
MacApple Silicon · Metal0.1.11WindowsWindows 10+ · unsigned installer0.1.5Linuxx86-64 · Vulkan + CPU fallback0.1.11Linux GPU support validated on NVIDIA RTX 2080 Ti and Intel integrated graphics. AMD validation is pending. Intel Mac downloads are paused.
Read the measurementsOpen source.
On your machines.
Potluck is open source. Here's why.
Potluck is now public under the MIT license. Why opening the code matters for the people running it, and why it helps us build better software.
Read the articleNotes from building and using Potluck.
More from the blogWhat is Potluck building next?
Company Brain connects current company knowledge to AI running on infrastructure you control, with reviewed outcomes guiding improvement. We’re starting with engineering investigations and handoffs. Read the plan.
Can my company use Company Brain today?
The integrated company workflow is in development. The current downloads provide local AI and existing machine-network features. Company permissions, reviewed improvement, and enterprise deployment controls are part of the build plan.
Is this for large organizations too?
Yes. The vision includes departments and large organizations. We’re starting with one engineering workflow and building the access controls and operational foundations needed for broader use.
Will our corrections automatically train a model?
The planned workflow keeps review explicit. Corrections can improve knowledge and evaluation cases first. Model adaptation would require approved examples, measured benefit, and controlled rollout.
What stays free, and what would companies pay for?
The open core is free for individuals and companies. The planned investigation, review, and knowledge-reuse workflow belongs in that core. We’re building additional company administration, monitoring, and support as a paid offering. Compare the plans.
Is Potluck open source?
Yes. Potluck’s core is public on GitHub under the MIT license, including commercial use. Future commercial tools will be identified separately. Model weights and third-party components keep their own licenses.
Is this one model running on several computers?
Yes. In a split setup, selected computers hold different parts of the same model and execute them together. That is different from sending the whole request to another computer that holds the entire model.
Can I try model splitting today?
Managed model splitting is experimental. Linux CLI 0.1.7 includes a CPU split engine. The tested Mac + Linux GPU setup uses development builds; it is not yet an end-to-end feature in the public desktop app. Every participant needs a compatible split runtime.
Can I just add up the RAM in my computers?
That is a starting point, not a fit guarantee. Each machine needs room for its operating system and working buffers. Model format, context length, GPU memory, and backend support all affect what can be loaded. Potluck checks the selected machines and reports where model weights actually land.
Will more computers make it faster?
Not necessarily. Splitting is a way to get more capacity. Speed also depends on how the model is divided, which parts run on CPUs or GPUs, and the connection between computers. More machines can add communication overhead.
Can I still use just one computer?
Yes. The desktop app can run a model locally on one computer. It is free to download and run locally, and requires an account. Connecting your own machines does not enroll them in a public contribution pool.
Help test
Company Brain.
Have a recurring investigation, procedure, or handoff in mind? Tell us about the work your team wants to improve. Compare the free core and planned company offering.
Open your email appOr write to hello@trypotluck.ai
What happens next
- 01
Tell us about the workflow. Describe the recurring work and the tools your team uses today.
- 02
Discuss the sources and success. We’ll talk about what information the work needs and what a useful result looks like.
- 03
Agree on a scoped test. When the workflow is ready, we can decide together what to try.
Available now: local AI on your machines.
Mac 0.1.11 · Linux 0.1.11 · Windows 0.1.5 · free to download and run locally.
Free to download and run locally. Account required.